DeepSeek Prompt Connector

The name is not subtle: this is a single node that hands your prompt to DeepSeek's chat API and gets a better one back. Type "cyberpunk street, rainy night" into it, and what comes out the other end is a paragraph someone actually tuned for an image model - subject, environment, lighting, lens, composition - ready to feed straight into your sampler's positive prompt.

Why would you want an LLM in the middle of your graph? Because prompting changed. On SDXL-lineage models a comma-separated tag list still works, but on the modern LLM-encoded stuff (Z-Image, Flux 2 Klein, the Qwen image family) your prompt is read by an actual language model, and it wants sentences and instructions, not (word:1.3) weights it will silently discard. Hand-writing that kind of prompt every time is tedious. This node is the shortcut, and it costs you nothing locally - it is a pure API call, so there's no model to download and no VRAM spent.

How it works

The mechanism is refreshingly small. The node builds a system prompt from a preset (or from your custom text), appends a "control block" to your input that tells DeepSeek which image model you're targeting, what style you want, and what language to output in, then POSTs it to https://api.deepseek.com/chat/completions with temperature and max_tokens. That's it - stdlib urllib, no openai package, no requests, nothing to install alongside it.

It returns two outputs, both plain strings:

• prompt - the final generated/improved prompt. This is the one you wire into your CLIP Text Encode node (or the LLM-encoder's text input).
• preview - the same prompt with a short header ([model: ...] [style: ...] [lang: ...]). Handy if you want to eyeball the settings you ran with, or pipe it into a label node; it's the same text, so don't wire both into a workflow that varies them independently.

The inputs that matter

You'll set most of these once and leave them alone, but a few deserve attention:

• api_key - your DeepSeek API key, pasted into the widget. Required; the node throws api_key is required if it's empty. Note that since it lives in the node, it also lives in any workflow JSON you save - don't share a workflow that still has the key in it.
• model - deepseek-chat (default) or deepseek-reasoner. Chat is the right call for rewriting prompts; reasoner is the R1-style thinking model, slower and pricier, and rarely worth it for this job.
• text (optional) - your input prompt or rough idea. Leave it empty and the node asks DeepSeek to invent a prompt from scratch, so it doubles as a random-idea generator.
• system_prompt_mode - six presets: improve, create from idea, photorealistic, cinematic, anime, or Custom. Pick Custom and you must fill custom_system_prompt, or it errors out.
• target_model - the one people misread. This does not select or download an image model. It's a hint to the LLM about which encoder your downstream node uses: pick sdxl and it writes SDXL-style descriptive tags; pick z-image turbo and it trims to concise, high-signal wording. The list is opinionated (yes, that's "Nano Banana" spelled weirdly), but the logic underneath is sound - different encoders genuinely want different prose.

Installing it

The README's stated path is just copy the folder in and restart, which is really a git clone in a hat:

cd ComfyUI/custom_nodes
git clone https://github.com/KaigorodovTuskul/ComfyUI-DeepSeek-API-connector

Restart ComfyUI and the node appears under text/deepseek as "DeepSeek Prompt Connector". You can also search "DeepSeek" in ComfyUI Manager if you'd rather do it from the UI. No models to fetch, no extra Python dependencies - the only thing you need that isn't in the repo is a DeepSeek account with a few credits on it.

Common gotchas

The obvious failure is a bad or missing key: you'll get a RuntimeError with an HTTP status (401 = wrong key, 402-ish = no credits). Make sure you've actually funded the account, since DeepSeek's API is pay-as-you-go and starts empty. The second classic is forgetting the text input and wondering why a blank node is inventing prompts - that's the feature. And one security note that applies to every LLM node in this category: it sends your text to DeepSeek's servers, and it stores your key in the workflow, so treat saved workflows with the key in them the same way you'd treat a password file.

It's a small pack with zero footprint in the wild, and the 0-impression search history says almost nobody's found it yet. But the mechanics are honest, the wiring is exactly one node wide, and for anyone running a Chinese-ecosystem image model who keeps hand-writing prompts, it's a genuinely handy little bridge.